Abstract: In order to solve the problem that more complex backgrounds such as weeds and shadows affect the accuracy of pear canopy image information extraction, a pear canopy image segmentation method based on improved DeepLabV3+ is proposed. This method introduces the attention mechanism into the backbone network of the DeepLabV3+ encoding part, between the hole space pyramid pooling module and the backbone network of the decoding part. After that, the important feature information will be paid attention to, which improves the model segmentation accuracy and ensures the segmentation efficiency. Taking the pear orchard with Yshaped trellis as the test object, the canopy segmentation experiment was carried out by collecting the canopy photos of pear trees by UAV. The results showed that the average intersection ratio, category average pixel accuracy and accuracy of the CBAMDeepLabV3+ model proposed in this paper for pear canopy image segmentation were 88.72%, 94.56% and 96.65%, respectively, and the segmentation time of a single image was 0.107 s. Compared with DeepLabV3+ and Se DeepLabV3+, the classification average pixel accuracy of CBAMDeepLabV3+ model for pear canopy segmentation was improved by 2.28% and 0.56%, respectively.

Key words: pear tree canopy, image segmentation, DeepLabV3+, attention mechanism, deep learning

摘要： 针对杂草和阴影等较复杂背景影响梨树冠层图像信息提取精度的问题，提出一种改进DeepLabV3+的梨树冠层图像分割方法。该方法将注意力机制引入到DeepLabV3+编码部分的主干网络与空洞空间金字塔池化模块之间和解码部分的主干网络之后，重要的特征信息将得到关注，提高模型分割精度的同时保证分割效率。以Y字形棚架梨园为试验对象，通过无人机采集梨树冠层照片，进行冠层分割试验。结果表明，提出的CBAMDeepLabV3+模型对梨树冠层图像分割的平均交并比、类别平均像素准确率和准确率分别为88.72%、94.56%和96.65%，分割单张图像时间为0.107 s。CBAMDeepLabV3+模型分割梨树冠层的类别平均像素准确率相比DeepLabV3+和SEDeepLabV3+分别提高2.28%和0.56%。

关键词: 梨树冠层, 图像分割, DeepLabV3+, 注意力机制, 深度学习